|
Background:
In the future,
machine agents will be able to communicate among themselves and
with the flexibility of human language. For example, robots will
be able to learn language and world understanding from direct
interaction with humans. Cognitive systems research focuses on
the development of natural and artificial information processing
systems (e.g. internet agents, adaptive agents, robots) capable
of perception, learning, decision making, communication and
action. They are designed to assist humans in a variety of
situations including everyday tasks, such as service/household
robotics, and highly-specialized situations, such as in
autonomous systems for defence.
Recent research in linguistics cognitive systems (Cangelosi et
al. 2005) has focused on the close integration of language and
other cognitive capabilities (i.e. integration of communication
with perception, categorization, action). This approach is based
on the important process of "grounding" the agent's lexicon
directly into its own internal representations. Agents learn to
name entities, individual and states whilst they interact with
the world and build sensorimotor representations of it. For
example, Steels (2003) studied the emergence of shared languages
in group of autonomous cognitive robotics that learn categories
of object shapes and colours. Cangelosi and collaborators
analysed the emergence of syntactic categories in lexicons
supporting navigation (Cangelosi 2001) and object manipulation
tasks (Marocco et al.2003; Hourdakis &Cangelosi 2005) in
populations of simulated agents and robots.
The use of this
grounded approach to the design of linguistic cognitive systems
is vital for overcoming the known difficulties in intelligent
agents whose linguistic abilities are purely based on abstracts
symbolic representations. This is the case of search engines
that only rely on text corpora and therefore cannot solve
lexical ambiguities that require consideration of contextual and
extra-linguistic knowledge. Grounded systems that have access to
the cognitive and sensorimotor representations of words can,
instead, succeed in solving these ambiguities. Equally
important, is the reverse: learning abstract categories and
situations, which are not directly observed in the world, can
only be grounded in language and communications among agents.
Current grounded agent and robotic approaches have their own
limitations, in particular for the scaling up of the agent's
lexicon since they can only use few tens of lexical entries (see
Steels 2003) and can deal with a limited set of syntactic
categories (e.g. nouns and verbs in Cangelosi 2001). This is
mostly due to the use of computational intelligent techniques
(e.g. neural networks, rule systems) subject to combinatorial
complexity (CC). The issue of scaling up and CC in cognitive
systems has been recently addressed by Perlovsky (2001). In
linguistic systems, CC refers to the hierarchical combinations
of bottom-up perceptual and linguistic signals and top-down
internal concept-models of objects, scenes and other complex
meanings. Perlovsky proposed the Modelling Field Theory (MFT) as
a new method for overcoming the exponential growth of CC in
computational intelligent techniques currently used in cognitive
system design. MFT uses fuzzy dynamic logic to avoid CC and
computes similarity measures between internal concept-models and
the perceptual and linguistic signals. More recently, Perlovsky
(2004) has suggested the use of MFT specifically to model
linguistic abilities. By using concept-models with multiple
senrorimotor modalities, a MFT system can integrate
language-specific signal with other internal cognitive
representations.
Perlovsky's proposal to apply MFT in the language domain is
highly consistent with the grounded approach to language
modelling discussed above. That is, both accounts are based on
the strict integration of language cognition. This permits the
design cognitive systems that are trully able to "understand"
the meaning of words being used be autonomously linking the
linguistic signals to the internal concept-models of the word
constructed during the sensorimotor interaction with the
environment. The combination of MFT systems with the grounded
agent simulations will permit the overcoming of CC problems
currently faced in grounded agent models and scale up the
lexicons in terms of high number of lexical entries and
syntactic categories.
The potential impact of this research for the development of
intelligent systems is great, also in the field of defence
interests. Cognitive systems are essential for integrated
multi-platform systems capable of sensing and communicating. In
future systems, robots and autonomous agents will be able to
learn language and world understanding from humans. In the area
of internet/text search engines the capability of truly
"understanding" the language query and corpora being used will
permit the design of more efficient search and data-mining
systems. In the area of intelligent agents for defence, the
design of cognitive systems able to develop autonomously their
own grounded lexicons will be beneficial in collaborative and
distributed tasks. (e.g. multi-agent exploration and navigation
in unknown terrains, etc.)
|
